Mnemolis Intents

A Home Assistant custom integration that exposes Mnemolis as a native LLM Tool API, making it available to any LLM-backed conversation agent (Ollama, OpenAI, Anthropic, etc.) directly from the HA UI.

Once installed, Mnemolis appears as a selectable API in your conversation agent options alongside the built-in Assist API. Your LLM can then search across your entire local knowledge stack with a single tool call.

What Mnemolis provides

• Offline knowledge — Wikipedia, Stack Exchange, iFixit, FreeCodeCamp, DevDocs via Kiwix
• Weather forecast — 3-day forecast via Open-Meteo (no API key required)
• News — Recent articles from your FreshRSS RSS feeds
• Web search — Live search via your local SearXNG instance
• Service status — Monitor status for all services via Uptime Kuma
• Home Assistant — Entity state summaries — lights, locks, sensors, motion, batteries, power

Available Tools

Tool	Description	Status
mnemolis	Routes queries to Kiwix, Open-Meteo, FreshRSS, SearXNG, Uptime Kuma, or Home Assistant	✅ Working
calculator	Evaluates math expressions, sqrt, trig, average	✅ Working
unit_converter	Converts between kitchen, weight, length, data, speed, and temperature units	✅ Working
calendar_day	Returns day of week and relative info for a given date	✅ Working

Requirements

• Mnemolis v3.2.0 or later running and reachable from Home Assistant
• Home Assistant 2024.6.0 or later
• An LLM conversation agent (Ollama, OpenAI, etc.) configured in Home Assistant

Installation

Via HACS (recommended)

1. Add this repository as a custom repository in HACS (type: Integration)
2. Install Mnemolis Intents
3. Restart Home Assistant

Manual

1. Copy custom_components/mnemolis_intents to your HA custom_components directory
2. Restart Home Assistant

Setup

1. Go to Settings → Devices & Services → Add Integration
2. Search for Mnemolis Intents
3. Enter your Mnemolis URL (e.g. http://192.168.3.5:8888)
4. Click Submit — HA will verify the connection before saving

Enabling for your conversation agent

1. Go to Settings → Devices & Services
2. Find your conversation agent (e.g. Ollama)
3. Click Configure
4. Under Control Home Assistant, enable Mnemolis
5. Save

Your LLM will now have access to the mnemolis tool and will use it automatically when answering questions that require looking things up.

Grounded answer synthesis (voice)

By default (v1.4.0+), the mnemolis tool asks Mnemolis to compose a short, grounded answer from the retrieved sources and relays that to your conversation agent — instead of the raw source blob (fused [SOURCE — DESCRIPTION] headers, article boilerplate, and headline excerpts). This is what keeps a voice satellite from reading section headers aloud: the answer comes back as a spoken-ready sentence, composed on your LLM backend under Mnemolis's own anti-hallucination gates (numeric grounding, source attribution, and an honest "the sources don't answer this" when they don't).

It's additive and reversible. On any synthesis skip, timeout, or gate rejection, Mnemolis returns no answer and the tool relays the raw result exactly as it did before — so turning this on can only replace boilerplate with a clean answer, never make replies worse.

Requires SYNTHESIS_ENABLED=true on the Mnemolis server (Mnemolis v3.55.0+). While that's off, the flag is a no-op and you get the raw result as before.

Two options are available under Settings → Devices & Services → Mnemolis Intents → Configure:

Option	Default
Grounded answer synthesis	On	Turn off to always relay the raw retrieved result
Answer length	voice	Default shape when the agent doesn't pick one per query: voice = ≤2 sentences for spoken replies; brief = one short paragraph; detailed = a few short paragraphs; digest = preserves many separate items

The agent chooses answer_style per query (v1.4.1+; guidance refined in v1.4.2), so a "list the headlines" request gets digest — preserving each item instead of collapsing ten headlines into one sentence — while a "summarize the news" request gets detailed, which weaves related items into one flowing summary. Ordinary single-answer questions stay at your configured default. The option above only sets that fallback default. digest requires Mnemolis v3.55.1+; the natural-prose detailed summaries are best on v3.55.2+.

For voice satellites, the recommended Mnemolis-side deployment note is LLM_KEEP_ALIVE=30m on a dedicated LLM backend so synthesis never pays a model load.

Calculator

Supports standard arithmetic, exponents, square roots, trig functions, and averages.

Examples:

• what is the square root of 1764
• what is average of 10, 20, 30
• what is sin of 45

Unit Converter

Supported unit pairs:

• Kitchen volume: cup, tablespoon, teaspoon, ml, pint, liter
• Weight: kg, lb, oz, g
• Length: km, mile, m, ft, inch, cm
• Data: kb, mb, gb, tb
• Speed: mph, kph, mps (meters per second)
• Temperature: use from_unit='c' to_unit='f' or from_unit='f' to_unit='c'

Amounts can be fractions: 1/2, 1 1/2, 0.75

Calendar Day

Ask what day of the week a date falls on, or how many days until an event.

Examples:

• what day is July 4th
• how many days until Christmas
• what day was January 1st 2000

Compatibility

Mnemolis Intents	Mnemolis
v1.4.2	v3.2.0 or later (digest requires v3.55.1+; natural-prose detailed summaries best on v3.55.2+)
v1.4.1	v3.2.0 or later (grounded answer synthesis requires v3.55.0+; the digest answer style requires v3.55.1+)
v1.4.0	v3.2.0 or later (grounded answer synthesis requires v3.55.0+ with SYNTHESIS_ENABLED)
v1.3.0	v3.2.0 or later
v1.2.0	v2.3.0 or later
v1.1.0	v2.0.0 or later

Part of the MiniNet stack

• Mnemolis — the knowledge broker backend
• MiniSense-T7S3 — ESP32-S3 room sensor node with voice assistant and CO2 monitoring